Accelerating Cartesian MRI by domain-transform manifold learning in phase-encoding direction.

Area of Science:

• Medical Imaging
• Machine Learning
• Biomedical Engineering

Background:

• Accelerating magnetic resonance imaging (MRI) acquisition is crucial for reducing scan times and improving patient comfort.
• Existing methods for reconstructing images from undersampled k-space data often face challenges with high-resolution datasets due to computational complexity.
• Deep learning approaches have shown promise in MRI reconstruction but require efficient model architectures.

Purpose of the Study:

• To develop a novel domain-transform framework for accelerated Cartesian MRI reconstruction.
• To significantly reduce the number of learnable parameters compared to existing manifold learning algorithms.
• To enable high-resolution MRI reconstruction with reduced computational and memory requirements.

Main Methods:

• Developed a domain-transform manifold learning framework (DOTA-MRI) with an initial analytic transform.
• Applied a 1D inverse Fourier transform (IFT) along the frequency-encoding (x-direction) of undersampled Cartesian k-space data.
• Utilized a symmetric deep neural network architecture with convolutional and fully connected layers for 1D global transform learning and image reconstruction.

Main Results:

• DOTA-MRI directly transforms undersampled k-space data into reconstructed images.
• The method drastically reduces learnable parameters from O(N^2) to O(N) compared to AUTOMAP.
• Achieved superior performance over nine other algorithms across various sampling ratios, datasets, and noise levels.

Conclusions:

• The DOTA-MRI framework offers an efficient and effective solution for accelerated high-resolution Cartesian MRI.
• The domain-transform approach significantly enhances computational efficiency and memory usage.
• Demonstrated the algorithm's generality and robustness for diverse MRI applications.